Finding and attribution
The default multimodal sender cache records image metadata before a prompt-length admission check, but a rejected request can prevent the engine receiver cache from receiving the payload. A later request that reuses the image hash is sent as a cache-only reference, and the receiver assertion can terminate the shared EngineCore.
Product: vLLM. AI system: GPT-5.5-Cyber. Team / organization: Patch the Planet (Trail of Bits + OpenAI collaboration) / vLLM maintainers.
CVE identifiers: CVE-2026-105753.
AI and human contribution (source account): The primary vLLM advisory identifies Patch the Planet (Trail of Bits + OpenAI collaboration) and states this vulnerability was discovered using GPT-5.5-Cyber. The advisory and linked PR describe this exact cache desynchronization; no autonomous or independent reproduction claim is made.
Human involvement is only described where the source account establishes it; unknown involvement does not mean none. Vendor confirmation of a flaw does not confirm every claim about AI contribution.
Discovery context: The vLLM security advisory provides a public disclosure date and describes the vulnerability. It does not report exploitation in the wild or establish whether the issue was unknown to defenders before discovery.
Remediation status: fix documented. The primary vLLM advisory lists 0.28 as the patched version boundary. It links a proposed fix; this record does not imply that any deployed instance has installed an update.
A source-documented fix does not establish installed patches or current safety. Source checking does not independently reproduce the finding or establish observed exploitation.
Disclosed 28 Sep 2026. AI attribution dated 28 Sep 2026. Sources checked 17 Nov 2026; these are separate dates, not a live installation check.
Recorded CVSS: Medium · highest recorded score 6.5. Documented technical effects: denial of service.
Severity, consequences and exploitation evidence
Case grouping uses the highest recorded CVSS score, with its version-specific severity band. It is descriptive, not a combined or editorial risk index. All recorded assessments remain below; scores from different versions and sources may disagree. CVSS specifications define version-specific scoring and severity bands.
- CVE-2026-105753: CVSS 3.1 · 6.5 (Medium) ·
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H. vllm-project/vllm GHSA-ph3r-5jfg-f84f · checked 17 Nov 2026.
CVE-2026-105753: denial of service.
Prerequisites / affected scope: An authenticated request against a vLLM deployment using the default mirrored LRU multimodal cache can trigger prompt rejection after rendering; a later request reusing that media hash can crash the shared engine. The advisory excludes processor_only, disabled-caching and shared-memory cache paths..
Exploitation evidence: The advisory describes a remotely reachable request sequence and a shared EngineCore crash; it does not report exploitation in the wild.. Impact source · checked 17 Nov 2026.
Documented fix: The primary vLLM advisory lists 0.28 as the patched version boundary. It links a proposed fix; this record does not imply that any deployed instance has installed an update.